Speech Models

Block unveils Dessa, a speech-to-speech model it says was trained on human conversation alone and runs locally on a MacBook Pro M4 Max Company claim

• Block says Dessa was trained "from scratch on nothing but human conversation. Sounds! No text, no pretrained models", and that as far as it knows Dessa is "the world's first AI you can have a realistic live conversation with that was trained only from human speech". Block says the model is full-duplex, so it listens and talks at the same time, that inference "runs faster than real time", and that "it runs locally on a MacBook Pro M4 Max or better". Block says it trained Dessa on a single NVIDIA GB200 NVL72.
• Block says Dessa hears and speaks emotion, sarcasm, laughter and a frustrated sigh natively, and that routing speech through text discards all of it. The Decoder reports Block gives curiosity, better human-AI communication and safety as its three reasons, the safety argument being that a voice system deaf to distress "just keeps going".
• Block has published only recorded demo clips and a waitlist — Dessa is not publicly available — and has posted an open call: "Calling all AI safety researchers. We're inviting a small group to explore Dessa, our model trained on speech alone, and help us understand what it can and can't do." One demo pairs Dessa with a third-party model, GPT-Live-1, under a disclaimer that its developer "is not affiliated with Block and has not reviewed or endorsed Dessa".
• Every capability claim here is Block's own; the company published no benchmark numbers and no independent evaluation exists. Block's product page carries no publication date, so this item is dated by the TechCrunch and Decoder reports of 24 October. Block names no safeguards against voice cloning or impersonation, and neither report cites an independent test of the locally-run claim.